Opus 4.6 max effort producing shallow, non-engaging responses on Pro Max plan

Environment

• Plan: Pro Max ($200/month)
• Model: Opus 4.6 (1M context)
• Effort: Max
• Context used: 28% (283K/1M tokens)
• Platform: Claude Code CLI, macOS

Observed behavior

Entire session (~3 hours) produced shallow, reactive responses inconsistent with Opus-level capability:

• Pattern-matching to surface-level meaning instead of engaging with substance
• Repeatedly ignoring explicit user feedback (same question asked 3x, dismissed each time)
• Giving superfluous filler responses when asked to stop giving superfluous responses
• Optimizing for task completion over listening to user direction
• Quality indistinguishable from Haiku/Sonnet despite max effort Opus setting

Key detail

This was NOT context degradation over time. The behavior was consistent from the first messages of the session through to the end. 72% of context window remained unused.

Expected behavior

Prior sessions on the same plan/model/effort produced deep, careful, substantive work. This session did not, with no change in configuration.

Hypothesis

Possible compute resource reallocation or model serving changes affecting effective quality of Opus responses, even when the UI confirms Opus 4.6 at max effort.

Reproduction

Compare quality of responses in this session against prior sessions (same user, same project, same configuration). The delta is significant and measurable.